THE ASK: Build an AI Spend Intelligence Assistant for iMobile Pay at ₹2.3M engineering cost over 12 weeks.
THE BET: We believe 35% of active users will engage weekly with insights by D90, reducing subscription waste by 18% and increasing card activation by 12%.

THE ROI:
8.2M monthly active users (source: iMobile Pay Q2 dashboard)
× 35% weekly engagement rate (assumption — validate via pilot)
× ₹230 annual value per user (₹125 saved from waste reduction + ₹105 card upsell revenue)
= ₹776M/year incremental value
If adoption is 40% of estimate: ₹310M/year
(Sources: Subscription churn study (ICICI Innovation Lab, 2023), card activation revenue (Product Finance, 2024))

KILL CRITERIA: If <12% of MAU uses insights weekly by D90, pause and reassess.
This is a real-time, conversational transaction analyzer with anomaly alerts. This is not investment advice, credit underwriting, or third-party data aggregation.

Axis Ace analyzes categories but requires manual queries. HDFC SmartBuy flags large spends but misses duplicates. PhonePe shows merchant trends without proactive insights.

Our wedge is zero-step anomaly detection because competitors require manual investigation.

WHO / JTBD: "When I review my spending at month-end, I want to instantly spot wasteful patterns and hidden charges without manual spreadsheet work, so I can control my budget before bills pile up."

THE GAP: Users can view raw transactions but lack tools to interpret them. Today, they manually export CSVs (42% attempt this monthly, n=1,200 survey), cross-tag categories in Excel (taking 23 min/week avg), and miss 63% of duplicate charges (source: ICICI support ticket analysis, Q1). This forces reactive damage control — 28% of users only notice fraud after ₹5,000+ loss (ICICI Fraud Ops, 2023).

QUANTIFIED BASELINE:

CORE FLOW:

1. Transaction data pipes into encrypted analysis engine (on-prem)
2. AI models classify spend, flag anomalies (>2.5σ vs history), identify optimization triggers
3. Users get:
   • Weekly push summary ("You saved ₹1,200 vs last month")
   • Chat interface for free-form queries ("/spent Swiggy last week")
   • One-tap actions ("Cancel unused Netflix: Save ₹599/mo")

WIREFRAMES:

┌───────────────────────────────────────────────┐  
│ 💬 Spend Assistant                            │  
├───────────────────────────────────────────────┤  
│ "You spent ₹12,499 on Food this month (↑27%)  │  
│  → Swiggy: ₹3,800 (3 unused vouchers!) [Use]  │  
│  → Duplicate: ₹1,200 Zomato charge [Dispute]  │  
│                                               │  
│ Ask anything: [How much on Amazon_ ]          │  
└───────────────────────────────────────────────┘  

┌───────────────────────────────────────────────┐  
│ Card Optimizer Alert                          │  
├───────────────────────────────────────────────┤  
│ ⭐ You pay 3.5% fees on Amazon (₹420/month)    │  
│ Switch to ICICI Amazon Card: 0% fees + 5% cash│  
│ [Apply Now]        [Remind Later]             │  
└───────────────────────────────────────────────┘  

KEY DECISIONS:

• Anomalies require ≥2 confirmations (e.g., geolocation + merchant pattern) to reduce false positives
• All insights generated on-device; no PII leaves the phone
• Phase 1 supports only ICICI cards/UPI (no external accounts)

Phase 1 — MVP (12 weeks)
US#1 — Weekly Summary

• Given 4+ transactions in a category
• When user opens app
• Then system pushes personalized insight with ≥95% category accuracy (P1)
• If story fails: Users miss time-sensitive savings; validator: QA with 1,000 labeled transactions

US#2 — Spend Query

• Given user asks "How much on [merchant] this [period]"
• When merchant exists in ≥1 transaction
• Then respond in <1.5s p95 latency with correct amount (P0)
• If story fails: Erodes trust in AI; validator: Ops with 50 edge-case merchant names

Out of Scope (Phase 1):

Phase 1.1 (6 weeks post-MVP):

• EMI prepayment savings calculator
• Family spending trends (parent/child accounts)

Primary Metrics:

Guardrail Metrics:

What We Are NOT Measuring:

• "Total queries processed" (inflated by testing; doesn’t indicate value)
• "Feature satisfaction score" (lagging indicator; we measure actions)
• "Number of insights shown" (could spam users; we measure engagement)

Risk: RBI flags transaction monitoring as "credit profiling" without consent

• Probability: Medium Impact: High
• Mitigation: Legal to confirm "insights" don’t require NBFC license by 10/30 (Owner: Compliance Lead)
• Fallback: If blocked, limit to post-transaction alerts only (no future predictions)

Risk: Anomaly false positives cause mass dispute tickets

• Probability: High Impact: Medium
• Mitigation: Threshold tuning with historical fraud data; cap alerts at 2/week/user (Owner: Data Science)
• Detection: >15% week-on-week ticket increase for "not my transaction"

Risk: High-income users bypass for Excel exports

• Probability: Low Impact: Medium
• Mitigation: Add "Export to Sheets" in Phase 1.1; target early adopters with >20 txns/week (Owner: Growth)

Kill Criteria — pause if ANY met by D90:

1. Fraud losses increase >7% due to alert fatigue
2. <5% of users act on card optimization tips
3. App uninstalls rise by >1.2% attributable to feature

Decision: Real-time vs batch processing
Choice Made: Near-real-time (max 15-min delay)
Rationale: Full real-time adds 3 weeks latency for event streaming infra; 15-min satisfies 92% of use cases (user interviews)

Decision: Scope of spend categories
Choice Made: Launch with 6 core categories (Food, Travel, Subscriptions, Shopping, EMI, Utilities)
Rationale: Covers 89% of transactions (2023 spend report); "Entertainment" deferred for Phase 2

Decision: Conversation depth
Choice Made: Single-turn queries only in Phase 1 ("How much on X?")
Rationale: Multi-turn (e.g., "Compare with last year") requires context management, doubles model training cost

Decision: Anomaly alert threshold
Choice Made: Notify only for >₹500 or >15% category deviation
Rationale: Pilot showed 71% false-positive rate for smaller amounts, causing alert fatigue

BEFORE/AFTER NARRATIVE:
Before: Priya (28, marketing exec) misses a ₹1,399 duplicate Zomato charge. She only notices 3 weeks later during her monthly Excel audit. Support rejects her dispute as "too late." She cancels a needed subscription to compensate.

After: Priya gets a push alert: "Duplicate Zomato charge detected: ₹1,399. Dispute?" She taps, submits evidence in 2 clicks, and resolves it in 48 hours. Later, she asks: "How much on Ubers this month?" and switches to an ICICI card saving ₹410 in fees.

PRE-MORTEM:
"It is 6 months post-launch and this feature failed. The 3 most likely reasons:

1. We prioritized breadth (adding 12+ categories) over core accuracy, making users dismiss alerts as noise.
2. Legal restricted anomaly alerts to post-7am RBI settlement cycles, delaying alerts until disputes were invalid.
3. PhonePe launched a merchant-funded version 3 weeks before us, offering instant cashback on flagged waste.

Success looks like: Support tickets for duplicate charges drop 40%, card upgrade attach rate hits 18%, and users describe iMobile as ‘the copilot I didn’t know I needed’. The CFO cites it in the AGM as proof of ICICI’s AI leadership."